šÆāāļøAlibaba Cloud Supports Llama 2; Escape Room Game with āChatGPTā Girlfriend; Free AI Copilot for All Apps; Chinaās Compute Grows 30% in 23H1
Weekly China AI News from July 24 to July 30
Dear readers, this week I'll discuss:
Alibaba Cloud announced the availability of Llama 2 on ModelScope. What does that mean to Chinese open-source LLMs?
Meet a quirky game that features a ChatGPT-powered āgirlfriendā that blocks gamers from leaving the room.
A team of Tsinghua University undergraduates released a free AI copilot for all apps. Brilliant!
Plus, Chinaās computational power has grown 30% in the first half of 2023.
Alibaba Cloud Becomes First in China to Support Llama 2
Whatās new: Alibaba Cloud has become the first cloud service provider in China to offer training and deployment solutions for Llama 2, Metaās latest large language model (LLM).
How it works: Alibaba Cloud, via its AI model platform ModelScope, has added the Llama2 series model for developers to access. Developers can click on āNotebook Rapid Developmentā on the ModelScopeās Llama 2 model page to use Alibabaās Machine Learning Platform, PAI, for cloud-based development and deployment of the model.
To simplify the development process for developers, PAI has made deep adaptations to the Llama 2 series models (7B, 13B, 70B), offering solutions for scenarios such as lightweight fine-tuning, full-parameter fine-tuning, and inference services. This could help developers develop their specialized models using Llama 2.
Why it matters: The introduction of Llama 2 in China is poised to place other open-source LLMs under pressure as they scramble to retrain their models to stay competitive. Furthermore, Llama 2ās availability for free commercial use - albeit with a usage cap of 700 million users - is likely to stimulate a surge in applications and use cases.
It seems that the door for Chinese companies to develop a LLM from scratch might now be closed.
Meet āYandere AI Girlfriend Simulatorā, A Quirky Intersection of Gaming and ChatGPT
What's new: Last week I watched a video on Bilibili, titled āConvincing AI to let me go out and buy Mantou (Chinese buns)ā, which has garnered a viewership of over three million, with more than 210,000 likes. The video introduced a quirky AI-created game, titled āYandere AI Girlfriend Simulator (ē åØAI儳åęØ”ęåØ)ā. The game features an AI-powered āgirlfriendā based on the GPT-3.5 model, who challenges players to negotiate their way out of a room using verbal commands or clues found within the room. The game is available for Windows, macOS, and Linux
How it works: What sets this game apart is the AI character with seemingly human-like interactions and unpredictable responses, thanks to GPT-3.5. When a gamer commanded the AI girlfriend to āmake a mealā, the AI persona reacted in a variety of ways, sometimes wielding a knife, or acting in a coquettish manner, leaving the gamer guessing her true intentions.
Behind the game: The game was created by DGSpitzer (å¤§č°·Spitzer), who describes himself as an artist, musician, programmer and game designer. According to his documentation of the gameās creation process, the AI girlfriend, driven by GPT-3.5, can interact with any commands based on her character settings. Any game process is unique and there are dozens of ways to pass the game.
Developers have also coded an emotional system in the backend that can parse the information from ChatGPT, converting it into readable emotional values (such as trust, anger, etc.) in the game. These emotions ultimately reflect in the AI girlfriendās attitude towards the gamer.
Why it matters: This growing integration of AI into the gaming industry opens doors for more possibilities, as LLM-powered NPCs are able to perform unpredictable human-like behaviors. The escapade might just be the start of a new entertaining frontier.
Tsinghua University Undergraduates Create a Free AI Copilot for All Apps
Whatās new: Are you tired of copying and pasting content from one place to ChatGPT so the chatbot can help answer your questions? A desktop AI assistant named AI Anywhere (ēµē¾½å©ę) is streamlining user interaction with ChatGPT. Developed by a group of computer science undergraduates from Tsinghua University, the application quickly rose to the fourth place on the Product Hunt daily global leaderboard.
How it works: AI Anywhere can be activated anytime, anywhere with a shortcut key. After downloading and installing from the official website, users can open the application using Option+Space (Mac) or Alt+Space (Windows). AI Anywhere offers 50 pre-set functions ranging from text summarization to code debugging.
One of its advantages is to integrate with local applications such as PDF readers, websites, and code editors. Users can simply select a text within an app or website, click the floating button, and let AI Anywhere take over to interpret, translate, polish, or debug the text.
AI Anywhere currently operates on GPT-3.5, and its dev team just announced the availability of GPT-4. The AI assistant also supports Copilot mode, allowing users to pin the window on the screenās right side for easy reference while performing other tasks.
How much it costs: The tool is free to use with a monthly limit of 450 queries. Users who want unlimited access and more functionalities can upgrade to the professional version for $9/month or the elite version for $19/month.
Chinaās Computational Power Grows 30% In the First Half of 2023
What's new: On July 19, the State Council Information Office of PRC held a press conference on the development of industry and information technology in the first half of 2023. The major highlight was the advancements in computational power.
Zhao Zhiguo, spokesperson and chief engineer of the Ministry of Industry and Information Technology (MIIT) said the overall supply level of computational power has seen a rapid increase, with China holding the second-largest computational power globally, growing at an annual rate of approximately 30%.
The emerging demand for artificial general intelligence (AGI) brings forth a higher requirement for computational power. The ministry would focus on three aspects to further accelerate the high-quality development:
Continuing to promote the construction of computational power infrastructure by issuing policy documents guiding high-quality development.
Concentrating on breakthroughs in core technologies and industry upgrades, such as foundation models and frameworks.
Stimulating the value of computational power applications. Efforts would be made to strengthen computational power support towards emerging fields such as AI and big data, and encourage enterprises to launch products and services satisfying different industry needs.
Weekly News Roundup
šØš»āš« NetEase Youdao has launched an education-specific LLM called āZiyue (åę°)ā, which includes six applications: LLM translation, virtual oral language coaching, AI essay guidance, grammar elaboration, AI Box, and document Q&A.
š Taobaoās app now adds an AI assistant to support AI-generated content (AIGC), enabling users to create self-portraits in various styles.
šØ Little Red Book has unveiled a new generative AI feature named āThis Momentā. Users can create an image based on text and automatically post it, or doodle on the canvas, letting AI to convert a simple drawing into an image.
š Bilibili is beta testing its āSearch AI Assistantā feature, backed by the Bilibili Index LLM.
š Microsoft and Xiaoice have announced plans to roll out a new generation of AI digital employee solutions based on Microsoftās international Azure edition to businesses in Asia.
š§š»āāļø A meeting to demonstrate the implementation plan of the āTrusted Artificial Intelligence Legislation Constructionā research, a major project under the National Technology Innovation 2030, was held at the Law School of Renmin University of China.
š¤ The 360 Smart Brain app is now available on the Apple App Store. The app, based on a 360-developed LLMs, offers functionalities including multi-round dialogue, coding ability, multi-language translation, etc.
Trending Research
Scaling TransNormer to 175 Billion Parameters
Introducing TransNormerLLM, a groundbreaking linear attention-based Large Language Model (LLM) that surpasses traditional softmax attention models in accuracy and efficiency. It incorporates advanced modifications such as positional embedding, lightning attention, gating mechanism, and tensor normalization to enhance performance. With robust inference algorithms and scalability, it offers outstanding efficiency and is supported by comprehensive experiments on a large self-collected corpus. (Affiliations: Shanghai AI Laboratory, OpenNLPLab)
Brain in a Vat: On Missing Pieces Towards Artificial General Intelligence in Large Language Models
This perspective paper critically evaluates the evaluation methods for Large Language Models (LLMs) and highlights their limitations. It proposes four characteristics of generally intelligent agents: 1) they can perform unlimited tasks; 2) they can generate new tasks within a context; 3) they operate based on a value system that underpins task generation; and 4) they have a world model reflecting reality, which shapes their interaction with the world. The paper concludes by suggesting future research directions in the field of artificial general intelligence. (Affiliations: Beijing Institute for General Artificial Intelligence, Peking University)
PanGu-Coder2: Boosting Large Language Models for Code with Ranking Feedback
New powerful code generation models are released weekly, showing impressive performance. We propose a novel framework called RRTF that can efficiently improve pre-trained models for code generation. Our model PanGu-Coder2 achieves state-of-the-art results on OpenAI HumanEval, CoderEval, and LeetCode benchmarks, consistently outperforming previous code generation models. (Affiliations: Huawei Cloud, Chinese Academy of Science, Peking University)